Parallel randomized load balancing: A lower bound for a more general model

Guy Even, Moti Medina*

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review


We extend the lower bound of Adler et al. (1998) [1] and Berenbrink et al. (1999) [2] for parallel randomized load balancing algorithms. The setting in these asynchronous and distributed algorithms is of n balls and n bins. The algorithms begin by each ball choosing d bins independently and uniformly at random. The balls and bins communicate to determine the assignment of each ball to a bin. The goal is to minimize the maximum load, i.e., the number of balls that are assigned to the same bin. In Adler et al. (1998) [1] and Berenbrink et al. (1999) [2], a lower bound of Ω(lognloglognr) is proved if the communication is limited to r rounds. Three assumptions appear in the proofs in Adler et al. (1998) [1] and Berenbrink et al. (1999) [2]: the topological assumption, random choices of confused balls, and symmetry. The topological assumption states that each ball's decision is based only on collisions between choices of balls. The confused ball assumption states that if a ball obtains the same topological information from all its chosen bins, then the ball commits to one of the chosen bins by flipping a fair coin. The symmetry assumption states that all the balls run identical algorithms, the same assumption holds for the bins. We extend the proof of the lower bound so that it holds without these three assumptions. This lower bound applies to every parallel randomized load balancing algorithm we are aware of (Adler et al., 1998 [1]; Berenbrink et al., 1999 [2]; Stemann, 1996 [3]; Even and Medina, 2009 [4]).

Original languageEnglish
Pages (from-to)2398-2408
Number of pages11
JournalTheoretical Computer Science
Issue number22
StatePublished - 13 May 2011


  • Balls and bins
  • Load balancing
  • Lower bounds
  • Static randomized parallel allocation


Dive into the research topics of 'Parallel randomized load balancing: A lower bound for a more general model'. Together they form a unique fingerprint.

Cite this